Your spam filter is L2-regularized logistic regression trained by gradient descent, and what you care about is a cost-weighted accuracy $a(\theta )$. A linear SVM on the same features scores higher on $a(\theta )$, but you have to ship the logistic model.

You evaluate your own training objective $J(\theta )$, the regularized log-likelihood you maximize, at the SVM's weights and find $J({\theta }_{\mathrm{SVM}})>J({\theta }_{\mathrm{LR}})$.